论文标题:主观轮廓感知修复模型及相关算法研究
Studies on Perceptual Completion Models and Algorithms for Subjective Contour
论文作者
论文导师 孙即祥,论文学位 博士,论文专业 电子科学与技术
论文单位 国防科学技术大学,点击次数 135,论文页数 132页File Size9914K
2006-10-01论文网 http://www.lw23.com/lunwen_88380972/
Subjective Contour; FACADE Theory; Contour Organization; Gap Completion; Tensor Voting; Orientation Estimation; Oriented Edge Detection
主观轮廓是一种在同质物理刺激的视野中根据图像特征的整体约束知觉到完整轮廓的视觉心理现象，体现了视觉系统惊人的感知识别和修复图像缺失部分的能力。主观轮廓作为一种理解目标轮廓知觉、形状知觉、深度知觉和视觉注意机制的重要现象，不仅仅有助于解决图像信息处理和机器视觉中的一些经典难题，而且对于认知科学具有特别重要的意义，在现代利用计算机进行图像场景分析和机器人视觉中具有广泛的应用前景。 本文主要对主观轮廓感知修复的模型和相关算法进行研究，从而为运用计算机模拟生物视觉对主观轮廓的感知识别过程奠定基础。论文主要取得了以下研究成果： 首先，提出了一个基于FACADE(Form And Color And Depth)理论的主观轮廓感知修复模型。该模型采用几何计算模型的建模思路建立了一个主观轮廓感知修复问题的综合性解决方案。在这一模型中，主观轮廓的感知修复过程被分解为轮廓特征提取、轮廓组织及缺口修复两个子过程。 其次，研究了轮廓特征提取的相关算法。本文建立的感知修复模型将主观轮廓的轮廓特征提取分为有向边缘检测和关键点检测两个部分。本文分析这两部分的共同点，采用取向估计这一关键技术对有向边缘检测和关键点检测进行了一体化的处理。由于取向估计是现有研究中的薄弱环节，本文对取向估计进行了比较系统的研究：给出了取向的数学定义；完善了取向估计的原理分析和相关证明；将lognormal滤波和非线性扩散滤波相结合提出了一种CLFND(Combined Lognormal Filter and Nonlinear Diffusion)算法，实验表明该算法与现有的几种典型的取向估计方法相比，具有更好的准确性和鲁棒性。在此基础上，本文进一步提出了基于取向估计的有向边缘检测算法和关键点检测算法。 最后，研究了轮廓组织及缺口修复的相关算法。本文主要完成了以下工作：一是针对张量投票方法处理主观轮廓时出现的问题进行了改进，并依据矩阵的特征值摄动理论分析了该方法误差产生的原因以及迭代对其影响；二是用图像处理方法代替FACADE理论中复杂的细胞动力学方程，提出了一种基于视觉竞争合作机制的轮廓组织及缺口修复算法，该方法通过DOG滤波器组、合作滤波器组和张量投票等方法分别模拟视觉系统中的简单细胞有向滤波、双极细胞合作滤波和终端截断后端点重组等主观轮廓形成过程；三是提出了一种基于可连接关系图的算法解决轮廓组织及缺口修复问题，该方法在轮廓特征提取的基础上，根据关键点的位置和取向信息判断其可连接关系，进而连接具有可连接关系但没有实际连接的关键点，最终完成轮廓缺口的修复。
The smooth completion of fragmented curve segments even when sufficientcontrast is lacking or in the presence of occlusions is an intrinsic skill of human visualsystem and subjective contour shows one of the compelling examples. Subjectivecontour can be defined as a visual psychological phenomenon which is to perceive aComplete contour in homogenious areas according to the implicit constrains presented inthe image. Recognizing subjective contour features is an intrinsic capability of humanvision, demonstrating the remarkable ability of the visual system. As an important cueto investigate the perception of contour,shape,depth and visual attention, research workon subjective contour can not only contribute to solve several problems in imageprocessing and computer vision, but also mean a lot to cognition science. The subjectivecontour theory will be proved to have many applications in modern life. This thesis addresses the research work on perceptual completion model ofsubjective contour and related algorithms, which will play a foundamental role forsimulating how the biovision system perceive subjective contours. The maincontributions of the thesis can be summaried as follows: Firstly, a perceptual completion model of subjective contour in the light ofFACADE (Form And Color And Depth) theory is presented. This model contains acomprehensive solution to solve the subjective contour problem, subdividing theprocess to contour feature extraction together with contour organization and contour gapcompletion. Secondly, the corresponding algorithms for contour feature extraction are studied:In the previous model, contour feature extraction consists of oriented edge detection andkey point detection. Through analyzing the similarities between these two parts, weadopt orientation estimation as the basic technology. Then a relatively systematic studyon orientation estimation is carried out: a mathematical definition for orientation isgiven; the guide theory for orientation estimation is extended and proved; we also putforward an orientation estimation algorithm called CLFND (Combined LognormalFilter and Nonlinear Diffusion) and demonstrate its efficiency in experiments.Furthermore, an algorithm for oriented edge detection and another algorithm for keypoint detection are brought forward, which are all based on orientation estimation. Thirdly, with regard to the contour organization and contour gap completion, thethesis presents three research results: one is an improved tensor voting algorithm forsubjective contour extraction. An iterative optimization procedure was developed basedon eigenvalue perturbation analysis to complete the contour gaps and to reduce theblurring effect of the original method. The second one is a computation approach for subjective contour which utilizes the cooperation-competition mechanism of biovisionsystem. The goal is to simplify the neural-computing of biovision process modeled inthe Boundary Contour System of FACADE theory, and to extract subjective contour in.a new way. The procedure consists of DOG filtering, cooperative filtering,post-processing of cooperative cues and regrouping of endpoints, corresponding to theoriented edge filtering by simple cells, cooperative filtering by bipole cells and end-stopeffect formulated by hypercomplex cells in biovision system respectively. The last oneis an algorithm based on connection relation graph, which judges whether two keypoints should be connected according to their positions and orientations and thenconnect the key points that Should be connected to complete the subjective contour.
|
